Peter Zijlstra wrote:
Hi,
This is yet another implementation of the PG_useonce cleanup spoken of
during the VM summit.
After getting bitten by rsync yet again, I guess it's time to insist
that this patch gets merged...
Andrew, could you merge this? Pretty please? ;)
Guys, this is a performance patch, right?
One which has no published performance testing results, right?
It would be somewhat odd to merge it under these circumstances.
And this applies to all of these
hey-this-is-cool-but-i-didnt-bother-testing-it MM patches which people
are
throwing around. This stuff is *hard*. It has a bad tendency to cause
nasty problems which only become known months after the code is merged.
I shouldn't have to describe all this, but
- Identify the workloads which it's supposed to improve, set up tests,
run tests, publish results.
- Identify the workloads which it's expected to damage, set up tests, run
tests, publish results.
- Identify workloads which aren't expected to be impacted, make a good
effort at demonstrating that they are not impacted.
- Perform stability/stress testing, publish results.
Writing the code is about 5% of the effort for this sort of thing.
Yes, we can toss it in the tree and see what happens. But it tends to be
the case that unless someone does targetted testing such as the above,
regressions simply aren't noticed for long periods of time. <wonders
which
schmuck gets to do the legwork when people report problems>
Just the (unchangelogged) changes to the
when-to-call-mark_page_accessed()
logic are a big deal. Probably these should be a separate patch -
separately changelogged, separately tested, separately justified.
Performance testing is *everything* for this sort of patch and afaict
none
has been done, so it's as if it hadn't been written, no?
-
